How it works
Every price comes from a page we can name. Below: how one price per task is worked out, and where each number is read from.
How a price per task is worked out
- List price.Take the tokens a benchmark run used. Price them at the maker's list price per million tokens.
- A seller's route.Price the same token mix at that seller's rates. Same run, different meter.
- A plan.Price the task at the plan's own rates, then multiply by the monthly fee and divide by the allowance. This assumes you use the whole allowance.
- Free routes.Never priced. They are listed as free, with their limits, and kept off the cost axis.
When a run reports only some of its tokens, the price is a range, not a point. Score-only benchmarks use a blend of 3 parts reading to 1 part writing.
What the price badges mean
- Listed
- The seller's own price page or API lists this price. A fresh read can confirm it; a listing is no promise that the service is available.
- Reported
- A catalogue such as models.dev or LiteLLM reports the price. We have not confirmed this number on the seller's own page.
- Estimated
- A plan's usable allowance comes from community measurements or a calculation, using the evidence recorded with the plan. The seller does not publish that dollar or token limit.
“Read” is the last live read of the source, not the day we parsed a cached file. A failed read or carried price stays stale. We warn after 30 days and say “too old” at 60. Unknown dates stay unknown. “Price since” counts days at the current observed price; it is not a promise that it will last. A recent cut compares a three-input-to-one-output token mix. Measurement dates on estimated plans describe the evidence, not a live read.
“Price unconfirmed” is a separate warning for a surprisingly cheap route, even when its seller lists it. A price under seven days old, or one we cannot confirm, gets a next steady price on the answer card when one is available at the same quality. A steady price is confirmed and has held for at least seven days. For imported usage, that comparison uses pay per use at the same token totals.
What a 60-day rule would leave out
This is a report only. All prices, picks and value lines keep the same routes.
Where the numbers come from
Status: ok means it was read; N warnings means it was read with caveats; failed means the last read errored.
What we don't do
- We don't use OpenRouter.
- Artificial Analysis data (artificialanalysis.ai) is licensed. It sits on a private page only, never on this one.
- Free tiers can end without notice. A free route here may be gone tomorrow.
- Plan limits are often estimates. Where a vendor does not publish its cap, we show "estimated limit" next to it. Published limits carry no such mark.